Neural Thickets and the Case Against Post-Training

When training large language models such as GPT or Gemini, the conventional view is that pretraining provides a general foundation, while additional stages—fine-tuning, reinforcement learning, and alignment training—are required to make the model effective at specific tasks. These post-training stages are often computationally expensive, time-consuming, and operationally complex.

But what if a substantial portion of that specialization already exists inside the pretrained model?

A recent paper from researchers at MIT explores exactly this possibility.

A useful analogy is to think of a pretrained model as a highly educated generalist. Traditional machine learning practice assumes that turning this generalist into a doctor, lawyer, engineer, or domain expert requires additional training through fine-tuning or reinforcement learning.

The MIT paper challenges this assumption. The researchers argue that many of these specialized capabilities may already exist within the pretrained model. Rather than creating new expertise through post-training, the process may simply be activating expertise that is already latent within the network.

They refer to this dense collection of latent capabilities as a neural thicket, a landscape containing many hidden experts embedded within the model's parameter space.

To explore this idea, the researchers introduced an approach called RandOp. Instead of performing gradient-based optimization through techniques such as backpropagation, PPO, or GRPO, RandOp perturbs model weights in numerous random directions. Each modified version of the model is then evaluated on a target task.

Conceptually, the process resembles adjusting a radio dial thousands of times, searching for frequencies that reveal stronger performance. The highest-performing variants are retained, and their outputs are aggregated through an ensemble voting process to produce a final prediction.

Formally, the method replaces traditional gradient optimization with a search procedure over random perturbations of the parameter space, where performance is determined empirically through evaluation rather than gradient updates.

What makes the result noteworthy is that the approach reportedly achieved performance competitive with established reinforcement learning methods such as PPO and GRPO, which are widely used in the post-training and alignment pipelines of frontier models.

The implication is not necessarily that reinforcement learning becomes obsolete. Rather, it suggests that large pretrained models may already contain a richer set of latent capabilities than previously assumed.

If this hypothesis continues to hold at larger scales, it could shift how researchers think about specialization in foundation models. Instead of investing significant compute into creating new capabilities through post-training, future systems may increasingly focus on discovering, activating, and combining capabilities that already exist within the pretrained network.

In that framework, post-training becomes less about teaching models new skills and more about finding the right experts hidden inside the neural thicket.